Social thread: 6 agent failure modes + Litmus testing
Engineers on X flagged common production agent failure modes — tool hallucinations, context overflow, silent runs, state loss — and recommended zero-dep debugging libs plus Litmus-style fault injection to validate circuit breakers. The posts call for replayable LLM call traces and injected faults to prove resilience in agent workflows. ( )
litmus-trace is published on PyPI as a record-and-replay helper for agent runs (package litmus-trace, version 0.1.1 on PyPI). (pypi.org) The litmus GitHub repository by rylinjam captures "every LLM and tool call" and advertises deterministic replay plus hooks for injected faults and reliability scoring to gate deploys. (github.com) OpenAI's Agents SDK documents a built‑in tracing layer that logs LLM generations, tool calls, handoffs and guardrail events and surfaces them in a Traces dashboard for per‑run debugging. (openai.github.io) An arXiv paper titled AgentRR (Agent Record & Replay) proposed in May 2025 formalizing the record‑and‑replay pattern for agents: record interaction traces, summarize into structured experiences, and deterministically replay to reproduce and guide behavior. (arxiv.org) LitmusChaos (the CNCF chaos platform) provides the execution primitives and ChaosExperiment/ChaosEngine CRs used to inject faults, revert them, and export post‑chaos metrics to Prometheus for steady‑state verification. (github.com) Latitude published a March 11, 2026 detection playbook that maps agent failure taxonomies to concrete observability gates, weekly reliability metrics, and regression‑gate practices for multi‑step workflows. (latitude.so) StateBase's public docs enumerate operational failure modes and a "7 Deadly Failures" framing intended to tie detection signals to automated recovery steps and observability instrumentation. (docs.statebase.org)
